Fast Loop Closure Detection Based on Holistic Features from SURF and ORB
LIU Guozhong1, HU Zhaozheng1,2
1. School of Information Engineering, Hebei University of Technology, Tianjin 300401, China;
2. ITS Research Center, Wuhan University of Technology, Wuhan 430063, China
摘要针对移动机器人SLAM(同时定位与地图创建)中的闭环检测问题,提出了一种基于SURF(加速鲁棒特征)和ORB(oriented FAST and rotated BRIEF)全局特征的快速闭环检测算法.首先利用SURF与ORB分别提取查询图像的全局特征,实现对图像的特征表征.在特征提取过程中,对查询图像进行归一化操作,并将归一化的图像中心直接作为SURF与ORB的特征点位置,从而避免了耗时的特征点定位过程.然后将归一化的图像直接作为特征点的邻域区域,把计算的SURF与ORB局部特征描述符作为图像的全局特征.为了融合SURF与ORB全局特征实现查询图像的位置识别,提出了H-KNN(混合K最近邻)的改进算法:WH-KNN(加权混合K最近邻).最后通过跟踪模型实现闭环检测,其核心思想是利用之前闭环检测的结果预测查询图像在地图图像中的位置范围.实验中分别使用采集数据集和牛津大学公开数据集对本文算法进行了验证,同时与传统的BOW(词袋)算法进行了对比.本文算法在两种数据集上分别达到了94.3%和94.5%的准确率,并且查询图像位置识别与全局特征提取的平均时间不到3ms.其准确性及计算速度都超过了BOW算法,可以准确快速地实现实时闭环检测.
Abstract:A fast loop closure detection algorithm based on holistic features from SURF (speeded-up robust feature) and ORB (oriented FAST and rotated BRIEF) is proposed for the loop closure detection problem in mobile robot SLAM (simultaneous localization and mapping). Firstly, holistic features of a query image are extracted for feature representation by using SURF and ORB. The query image is normalized and its center is directly set as the feature point position of SURF and ORB in the process of extracting holistic features, so that the time-consuming step of feature point localization is avoided. Afterwards, the normalized image is directly used as the patch of the feature point. The local feature descriptors computed from SURF and ORB are then used as the holistic features of the image. In order to fuse the holistic features from SURF and ORB for image matching, an improved method of H-KNN (hybrid K-nearest neighbor) called WH-KNN (weighted hybrid K-nearest neighbor) is also proposed. Finally, loop closure detection is finished by a tracking model. The core idea behind the model is to utilize the previous loop closure detection results to predict the position of the query image in the map. In the experiments, the proposed algorithm is not only tested with the collected datasets and the public datasets of Oxford University, but also compared with the classic BOW (Bag of Words) algorithm. It can achieve 94.3% and 94.5% detection accuracy on these two datasets, respectively. In average, it takes less than 3 ms to match each query image and extract holistic features. The test results demonstrate that the proposed algorithm outperforms the BOW method both in term of detection accuracy and computation efficiency. Therefore, the algorithm is accurate and fast enough for real-time loop closure detection.
[1] Thrun S, Leonard J J. Simultaneous localization and mapping[M]//Springer Handbook of Robotics. Berlin, Germany:Springer-Verlag, 2008:871-889.
[2] Xu H X, Zhou W, Zhu J. 3D visual SLAM with a time-of-flight camera[C]//IEEE International Workshop on Signal Processing Systems. Piscataway, USA:IEEE, 2015.
[3] 付梦印,吕宪伟,刘彤,等.基于RGB-D数据的实时SLAM算法[J].机器人,2015,37(6):683-692.Fu M Y, Lü X W, Liu T, et al. Real-time SLAM algorithm based on RGB-D data[J]. Robot, 2015, 37(6):683-692.
[4] Williams B, Cummins M, Neira J, et al. A comparison of loop closing techniques in monocular SLAM[J]. Robotics and Autonomous Systems, 2009, 57(12):1188-1197.
[5] Lategahn H, Stiller C. Vision-only localization[J]. IEEE Transactions on Intelligent Transportation Systems, 2014, 15(3):1246-1257.
[6] Clemente L, Davison A J, Reid I D, et al. Mapping large loops with a single hand-held camera[C]//Robotics:Science and Systems III. Cambridge, USA:MIT Press, 2008:27-30.
[7] Williams B, Cummins M, Neira J, et al. An image-to-map loop closing method for monocular SLAM[C]//IEEE/RSJ International Conference on Intelligent Robots and Systems. Piscataway, USA:IEEE, 2008:2053-2059.
[8] Cummins M, Newman P. FAB-MAP:Probabilistic localization and mapping in the space of appearance[J]. International Journal of Robotics Research, 2008, 27(6):647-665.
[9] Paul R, Newman P. FAB-MAP 3D:Topological mapping with spatial and visual appearance[C]//IEEE International Conference on Robotics and Automation. Piscataway, USA:IEEE, 2010:2649-2656.
[10] Kejriwal N, Kumar S, Shibata T. High performance loop closure detection using bag of word pairs[J]. Robotics and Autonomous Systems, 2016, 77:55-65.
[11] 李博,杨丹,邓林.移动机器人闭环检测的视觉字典树金字塔TF-IDF得分匹配方法[J].自动化学报,2011,37(6):665-673.Li B, Yang D, Deng L. Visual vocabulary tree with pyramid TF-IDF scoring match scheme for loop closure detection[J]. Acta Automatica Sinica, 2011, 37(6):665-673.
[12] 李永峰,张国良,王蜂,等.一种基于历史模型集的改进闭环检测算法[J].机器人,2015,37(6):663-673.Li Y F, Zhang G L, Wang F, et al. An improved loop closure detection algorithm based on historical model set[J]. Robot, 2015, 37(6):663-673.
[13] 李维鹏,张国良,姚二亮,等.基于空间位置不确定性约束的改进闭环检测算法[J].机器人,2016,38(3):301-310, 321.Li W P, Zhang G L, Yao E L, et al. An improved loop closure detection algorithm based on the constraint from space position uncertainty[J]. Robot, 2016, 38(3):301-310, 321.
[14] Ma H, Chan J C, Saha T K, et al. Pattern recognition techniques and their applications for automatic classification of artificial partial discharge sources[J]. IEEE Transactions on Dielectrics and Electrical Insulation, 2013, 20(2):468-478.
[15] Li F F, Fergus R, Perona P. Learning generative visual models from few training examples:An incremental Bayesian approach tested on 101 object categories[J]. Computer Vision and Image Understanding, 2007, 106(1):59-70.
[16] 梁志伟,陈燕燕,朱松豪,等.基于视觉词典的单目视觉闭环检测算法[J].模式识别与人工智能,2013,26(6):561-570.Liang Z W, Chen Y Y, Zhu S H, et al. Loop closure detection algorithm based on monocular vision using visual dictionary[J]. Pattern Recognition and Artificial Intelligence, 2013, 26(6):561-570.
[17] Lowe D. Distinctive image features from scale-invariant keypoints[J]. International Journal of Computer Vision, 2004, 60(2):91-110.
[18] Bay H, Tuytelaars T, van Gool L. SURF:Speeded up robust features[M]//Lecture Notes in Computer Science, vol.3951. Berlin, Germany:Springer, 2006:404-417.
[19] Rublee E, Rabaud V, Konolige K, et al. ORB:An efficient alternative to SIFT or SURF[C]//IEEE International Conference on Computer Vision. Piscataway, USA:IEEE, 2011:2564-2571.
[20] 龙超,韩波,张宇.基于Kinect和视觉词典的三维SLAM[J].计算机应用,2016,36(3):774-778.Long C, Han B, Zhang Y. Three-dimensional SLAM using Kinect and visual dictionary[J]. Journal of Computer Applications, 2016, 36(3):774-778.
[21] Mur-Artal R, Montiel J M M, Tardos J D. ORB-SLAM:A versatile and accurate monocular SLAM system[J]. IEEE Transactions on Robotics, 2015, 31(5):1147-1163.
[22] Hu Z Z, Hu Y Z, Li Y C. Fast road sign recognition based on ORB-encoded holistic and local features[C]//95th Annual Meeting of the Transportation Research Board on Artificial Intelligence and Advanced Computing Applications. 2016.
[23] Oxford Mobile Robotics Group. New college and center city dataset[DB/OL].[2016-06-28]. http://www.robots.ox.ac.uk/~mobile/IJRR_2008_Dataset/data.html.